LLM abstraction levels inspired by fish eye lens

What is this voodoo?? This is the among the loveliest sites I've read in a long time! <3

Beautifully designed, executed and presented. With just enough tension to pull the reader in...

I had to look up the author (who I was not familiar with, tbh) and her other work. Good job!

Beautifully done. 12/10, would read again!

My cursor got schooled.

Intriguing example with the shoemaker story, but initially confusing. After playing with it, okay, but still perhaps more confusing than the concept which is clear.

The trick with the cursor suggests: what if the context was ambient, like a graph of concept bubbles around whatever paragraph has focus? Or, a step further, as if the paragraph was a card in an array of cards swiped across a desk in "is this your card?" style, with fisheye focus on paragraph text, past and future paragraphs as cards to left and right of this one, with summarized concepts bracketing those from above and below?

In any case, this is onto something, and worth experimenting with various UIs.

I have no idea who the author is but that was the most beautiful blog Post I’ve ever come across. I would be perturbed if didn’t know ai was used to help make the fish but I’m not trying to detract from my praise. Good post op

Unless someone starts paying people to read a system like this will probably be inevitable in our information saturated age were nobody has the time to read anymore. Pulling in decades or centuries worth of context and motivation for the passage you are reading is definitely a worthwhile goal.

I really like the layout of the reimagined wikipedia it has the feeling that there is an entire group of people and subjects the care about and rely on the central topic. It's the exact same feeling as studying at a good school. The narrative that this circle creates can be a powerful motivator for learning. With wikipedia and other wikis, most articles feel like your reading an expense report and 99% of links are equivalent to the "general topics" and seem to serve more as distractions from your topic rather than a circle around it.

An outliner would be a good way to represent this info: something like Workflowy, Org mode, Ravel or Dynalist. You get nested bullet points that you can collapse.

Workflowy: https://workflowy.com/ Ravel: https://ravel.acenturyandabit.xyz/

A term I've heard since the ...2000's?...is "focus+context" [1], often used in visualisation. I'm looking for a very similar UI construct right now, for balancing simple high-level terms for non-technical people but digging in for people who want more detail, or who want a definition. My thinking is to have a sidebar with more info, but a closable popup might work better for mobile. Don't like popups but it's e.g. how iOS does "look up" when focusing on a word.

[1] And article hints at it with the lines "A fish eye lens doesn’t ask us to choose between focus and context" and "This concept isn’t new—it’s foundational to fields like data visualization."

Another article with a similar concept, around AI, but not LLMs: https://blog.jtoy.net/choosing-the-appropriate-abstraction-l...

The biggest problem with complexity is that you have it. I applaud wanting to manage complexity, but there's a difference between managing complexity and simply hiding it.

You want the right information, summarized correctly, from the right source, to the right person, and just in time. We can barely, sometimes, do that if we're lucky with deterministic systems... We loose most of the control of any of these pieces in probabilistic systems. Maybe that works for you, but it feels like hubris to use it anywhere I might actually think I could use it.

I think the process of learning is in fact making larger pictures into smaller quotes and this does not help learning in the way that it’s trying to. The process of diluting context is something that fundamentally must happen inside a person and trying to replace that with Ai is contrary to the point.

I do find significant use for this in something like review and considering the legibility of my own note taking, that is one handsome solution.

How do I disable the fish following my mouse? I skipped the content because of that. Too, damn, distracting.

I think this is a great way to explain the idea of a knowledge graph. There's probably a whole field of ux research opening up to interact with knowledge graphs and using llms to create them, parse them and display them to users

This sounds uncannily similar to Barbara Minto's "Pyramid Principle"…

Isn't this the original design of hypertext as intended by Ted Nelson? Like, instead of clicking links and navigating to somewhere else, the relevant information would actually accompany the text?

Beautiful blog post!

I've built something similar and have yet to find a compelling use case for it. In my mind it seemed like such an obvious idea that I thought I'd start using it instead of Wikipedia. However, in practice I seldom care about the broader or more specific topics for the thing I'm looking into. It's strange though because I'm a compulsive rabbit holer but it just doesn't trigger the same urge to explore related nodes as Wikipedia does.

With that said, this UI is much slicker so maybe that is the missing piece.

This is such a lovely article and I relate to the idea as I have thought of something similar when I was building a passion project (https://github.com/pixlie/gitplay).

The idea was to use the Git commit log to show how a software evolved into what it is today. I needed to think of zooming in and out of context, where there are thousands of commits involved. I am not a good UI person so I will definitely take inspiration from this.

I created something related long ago, when there weren't any LLMs. I do think it works better when human authors can choose when to add information and how:

https://skorokithakis.github.io/expounder/

Haystack Editor uses a UI of inter-function control flow graphs, but lacks the ability to "summarize" surrounding code. (In my usage I found it was buggy, and useful for tracing a specific level of knotty code, but had issues relative to traditional IDE tabs/panes for both small files and flipping between multiple "aspects" of a system.)

The idea is innovative and quite interesting, but once I'm interacting with the demos and imagining I'm interacting with text this way, I find it quite distracting and hard to navigate between text content, since almost every piece of text is constantly changing while I'm switching focus.

Oh dear, someone has invented an even more compelling way for me to get stuck in a wikipedia article reading hole.

The example is incorrect. On the paragraph where the elves (dwarves??) get the clothes, it is specified no cut pieces are left out for work, but the subsequent summary is the elves settle to work?

I don't think this is a particularly good showcase.

we've had other expressions that covered this concept as well: 30,000' view, boots on the ground, deep in the trenches. Forest through the trees.

It's great that someone has a new saying, but it's still the same concept. It's not like we've just been shown a new wheel. It's just been cleaned up and given fresh coat of armor all, and maybe chrome rims instead of the ones from the factory

As long as I can turn it off, I get distracted so easily. I can barely code full-screen with one function visible at a time.

That wikipedia view sounds terrifying.

Those who don't know the pyramid principle are doomed to reinvent it.

Reminds me of hyperbolic deep learning.

Great explanation! It's actually a kind of version of "TL;DR," which stands for "Too long, don't read." One of AI's benefits is personalization, which means that you can jump between abstraction levels, or in other words, information density levels.

looks amazing.